Real-time voice channel defence

One model.
Three attacks.
It runs where you are.

Cloned voice, vishing script and social engineering — one decision engine, on-premise, inside your own network. No datacentre required.

Book a 15-minute call Watch two real calls →
Shipped detector · public benchmark
3.05% Equal Error Rate — ASVspoof 5 dev track_1
AUC · 0.9943
N · 20,000
MODEL · v11_doomsday
MEASURED · 2026-07-21
Live verified · v11.detector.eer
Delivery modelOn-premise, customer-controlled
PositionPassive on the call
Runs today in our own development and demo environment on NVIDIA Triton. No customer production deployment.
01
The problem

External sources
Register-backed

The banks are already paying for it.

Not a forecast — the reimbursement line regulators already publish.

£316M reimbursed by UK banks since mandatory APP reimbursement began
SOURCE · UK Payment Systems Regulator
PERIOD · 2024-10-07 → 2026-03-31
Live verified
88% reimbursement rate on in-scope claims — up from 61%
SOURCE · PSR APP dashboard
PRIOR · UK Finance 2024
Live verified
57% of adults worldwide hit by a scam attempt in twelve months
SOURCE · GASA 2025
N · 46,000 adults · 42 markets
Live verified

A perfect clone passes the match. So we ask the other question.

Voice biometrics asks who is this. When the clone is good enough the answer is correct — and wrong. VoiceGuard asks whether a human is speaking at all, and whether the conversation is behaving like a scam.

02
The evidence

Deterministic render
from raw capture

Two real calls. Nothing staged.

One cloned voice, one real human whose conversation is the problem. Every number on screen is computed at playback from the raw capture.

3:01
The long version, with every number and how it was measured — 4:02
Repo verified · capture 2026-08-09
03
VoiceGuard

Shipped · measured

The measurement, with its scope written next to it.

3.05% Equal Error Rate on the public ASVspoof 5 benchmark
AUC · 0.9943
CORPUS · ASVspoof 5 dev track_1
N · 20,000
MODEL · v11_doomsday
MEASURED · 2026-07-21
REGISTER · v11.detector.eer / .auc / .corpus
Live verified
15.2% the challenge's own baseline, on the same set

Same corpus, same split. The comparison is the point — not the number on its own.

What this number does not cover. The benchmark is codec-clean. Lossy telephone-band compression is measured separately and is exactly what a bank pilot calibrates. We would rather you heard that from us.

Repo verified · scope limit stated
Try the Audio Challenge VoiceGuard technical deep dive — in preparation
D1 · Where we attach

The call path is never interrupted.

CALLERBANK VOICEGUARD OBSERVES · DOES NOT CARRY

Nothing of ours sits in the media path. If we fail, the call does not. In-app and conference-join capture today; carrier-side SIPREC/SBC integration is the enterprise-scale target.

SIPREC/SBC · docs / roadmap — not shipped
D2
Verdict path

Mechanism only
Threshold withheld

It does not shout on one suspicious moment.

0102030405 audio segmentparallel streams fusionagreement acrosssuccessive segments verdict STREAMINGKAN ENSEMBLE ONE VERDICT PER SEGMENTTEMPORAL RULEREPORTED

The system does not report on a single suspicious moment: several consecutive segments have to agree before it speaks. The exact number of segments is deliberately not published — it would tell an attacker what to stay under.

Threshold value · withheld by design
04
Commercial

Capacity, not usage

Priced on capacity, so full coverage is never the expensive option.

An annual licence on concurrent channel capacity — not per minute, not per call.

Per-minute pricing forces a bank to sample, and sampling means not looking exactly where the fraud comes in. It also requires metering that phones home — which is at odds with an on-premise, customer-controlled delivery model.

Second product lineHYDRA NDR

The attacks that have no signature yet.

Unsupervised KAN models trained on benign traffic only, combined with deterministic glass-box rules. Every decision traces back to the signal that triggered it.

D3 · Detection surface Named gaps — not omissions
Command-and-control channelsDemonstrated
Scanning & reconnaissanceDemonstrated
Junk-data C2Not on this surface
Lateral movementDemonstrated
Protocol abuse & impersonationDemonstrated
Steganographic C2Not on this surface
ExfiltrationDemonstrated
DNS-layer evasionDemonstrated
NAT traversalPacket-header tier

The three on the right are not demonstrated from the current network-observation surface, and we say so rather than counting them. A vendor that claims everything is claiming nothing.

Repo verified · per-technique evidence records
Book a 15-minute call HYDRA technical deep dive →
05
The horizon

Same engine
Different maturity

Where this goes — and what is not built.

VideoGuard Video deepfake detection. Adding the voice stream to the video model measurably improves the hardest cell. Internal milestone, not externally validated. Prototype · measured
TextGuard Messaging and SMS. Designed, not written. Not built
LinkGuard Links, pages and QR codes checked before they open. Designed, not written. Not built

Witness is a capability of VoiceGuard, not a separate product — and it reports three states, never two: verified, flagged, and not verifiable. The grey state is mandatory; without it the system would be claiming that unsigned means fraud.

06
Company

Engineering-led. Evidence first.

Work began in 2025. BluewaveAI Kft. was registered on 13 February 2026 in Budapest. Every claim above carries its own provenance marker; nothing on this page is presented at a maturity it has not reached.

IBM — ESA-qualified · Partner PlusPartner status
HPE Partner Ready VantagePartner status · HYDRA PoC test plan v1.3 submitted 2026-05-29 · execution pending
University of Szeged (SZTE)Research partnership · in progress
NVIDIA InceptionProgramme member
居合
Our discipline

The blade is not drawn in anger. It is drawn in readiness.

Iaido trains for the moment you hope never comes. Thousands of repetitions — not to fight, but so that when it arrives the response is already there. We do not react to threats. We are ready before they have a name.

Junbi · Preparation
Seido · Precision
Sokuō · Immediacy

Fifteen minutes, and we scope it against your channel.

No deck. We ask what your voice channel looks like, and tell you where we would and would not help.

Book a 15-minute call Try the Audio Challenge →